Claude for Accounting & Finance: What Actually Works (2026)

In October last year I ended up in Business Insider for something the profession does not usually get profiled for — teaching working accountants to build small applications by prompting AI. Twelve months on, the more useful story has shifted. It is no longer "can an accountant learn to code with AI" but "can an accountant do their job faster with Claude by their side." And on that second question, the answer in 2026 is clearly yes, with caveats worth writing down.

This piece is what I share with audit partners, finance controllers, and CFOs when they ask how to start — and where to be careful. It is opinionated. It draws on what Anthropic itself publishes, on live deployments at Big Four firms, and on what I see practitioners do inside YYC Advisors and across the Malaysian Institute of Accountants (MIA) board rooms.

1. Where Claude is genuinely useful today

Anthropic publishes a Claude for Financial Services product line, and the honest description is closer to "a very good associate" than "an autonomous accountant." Five workstreams show up over and over in real deployments:

• Month-end close. Drafting variance commentary against last month, last year, and budget. Draft goes to a reviewer; reviewer edits down.
• Reconciliation. Reading two exports (bank vs GL, subledger vs GL, AR aging vs sales) and surfacing the differences with a plausible explanation for each.
• Audit prep. Reading a set of contracts or a policy document, extracting the audit-relevant clauses, mapping them to MFRS or ISA requirements, and drafting the working paper. Human confirms the mapping.
• Tax memos. Drafting first-pass tax positions with the reasoning shown — useful for RPGT computations, group relief eligibility, transfer pricing rationale. Every claim gets checked by a chartered accountant before it lands anywhere client-facing.
• Board reporting. Turning a P&L export into a one-page executive summary with the numbers the board will actually ask about surfaced first.

The CFO.com report from 2026 lists deployments at JPMorganChase, Goldman Sachs, Citi, AIG, and Visa — and separately, PwC has committed to training 30,000 U.S. professionals on Claude and integrating it into their audit and Office-of-the-CFO offerings.

2. The internal Anthropic finance story that matters

Anthropic's own finance team publishes a lot about how it uses Claude on itself. Two data points worth the read. During a NetSuite-to-Workday migration, they used Claude to validate historical data in batch — work that would have taken hundreds of manual hours reduced to roughly 20 seconds per year of data. Their finance team now runs on 150+ Claude Skills — small packaged prompts that automate the recurring bits of month-end, forecasting, and reporting.

The lesson for Malaysian finance teams is not to copy 150 skills wholesale. It is to notice that the practitioners who get the most out of Claude are the ones who invest in packaging the parts of their job that are repetitive. Nobody builds a good AI-augmented finance function by using Claude ad hoc every day.

3. The 5 workstreams where Claude changes accounting

Below is the map I share with heads of finance and audit partners when they ask what to prioritise. Five workstreams, each with a "what changes" note, in the order I would sequence them.

4. What Claude cannot yet do

The failure modes are worth stating plainly.

• Arithmetic on messy spreadsheets. Claude reads spreadsheets well now — but numeric operations on large sheets should still route through code or a proper spreadsheet engine, not free-form reasoning. If it is a formula, ask Claude to write the formula, not compute the answer.
• Regulator-facing filings. Do not put Claude's output into an MFRS-audited financial statement, a SST-02 filing, or an LHDN e-Filing form without a chartered accountant reviewing every number and every line. The auditor and the tax agent's professional obligation does not disappear because you used AI to draft.
• Fraud detection. Claude can flag anomalies in transaction data — unusual counterparties, round-number payments, sudden vendor changes. It cannot conclude fraud has occurred. It also cannot cross-check bank records or SSM filings without those being provided as inputs.
• Client confidentiality. Claude Free and Pro plans process customer data to improve the model unless explicit privacy settings are toggled. Business and Enterprise plans do not. If you are handling client accounts, use Claude for Business or Enterprise, and confirm the data-processing agreement in writing.

5. The Malaysian audit context

The MIA has been publicly supportive of AI use, provided the professional obligation stays with the accountant. Practically, this means:

• Every AI-assisted working paper needs a human reviewer's signature on the audit file.
• Client engagement letters should disclose that AI tools may be used to assist — you do not need consent for use, but you should be transparent that it happens.
• Confidential audit information must not flow through consumer AI accounts. Firm-issued Business/Enterprise accounts are the minimum bar.
• Working papers must record which portions were AI-drafted, so the file survives a peer review or a regulatory inspection.

I have written elsewhere on AI adoption paths for Malaysian accountants and on AI governance frameworks for regulated Malaysian environments. Both are longer treatments of the compliance side.

6. How to start — a partner's reading

If you are a Managing Partner or Head of Finance reading this, the practical order I recommend is this. First, buy a Claude Team or Business plan and roll it out to the finance team, not the whole firm. Second, pick one recurring workflow — usually variance commentary in the monthly close pack — and prompt-engineer it until the reviewer edits less than 20% of the draft. Third, package that prompt as a Skill or a shared prompt library so it works the same way every close. Fourth, and only then, expand to a second workflow. Fifth, formalise the review and audit-trail policy before touching anything client-facing.

Firms that skip step three usually plateau at "a couple of partners use Claude for their own stuff." That is fine, but it does not change the economics of the firm.

7. HRDC funding for AI in accounting

Malaysian employers can claim training under the HRD Corp SBL-KHAS scheme for structured AI upskilling programmes. AITraining2U's AI Automation programme and Claude training curriculum are HRDC SBL-KHAS claimable for eligible Malaysian employers. Finance teams typically start with the automation curriculum and layer Claude-specific practice in the second half.

Two years ago I told accountants they had to learn to prompt AI to stay relevant. In 2026 the ones who did are running smaller, higher-margin practices. The ones who did not are the ones asking me to teach them what I taught the first cohort. The training window is closing.